ResearchPublished All news from that day
Microsoft Research reports its TokenEconomizer plugin cut coding-agent cost per task by 18.4% to 30.9% on SWE-Bench
The figures are the researchers' own, from a 16 November Microsoft Research post on a paper from UC Berkeley, Microsoft Research and Microsoft Security.
Check our sources · 7 facts from 1 sourceMicrosoft Research reports that TokenEconomizer, an open-source plugin that learns from a coding agent's past sessions in a repository, cut mean cost per task by 18.4% to 30.9% on SWE-Bench. The result covers 277 tasks in six repositories and five agent-model combinations, and pass@1 rose in four of the five, the researchers say.
The post names Claude Code as the only agent; it does not identify the other four combinations. Profiling Claude Code on 60 SWE-Bench tasks, the researchers found that 30 percent of tool calls repeated an earlier action in the same repository, and that 16 of 23 tool descriptions sent per turn were never called.
On Web-Bench they report costs 16.0% to 21.2% lower, and on security auditing 6.4% to 19.6% lower. They say lowering reasoning effort saved similar or larger amounts but substantially degraded performance. The code is on GitHub at microsoft/TokenEconomizer, and the post says the plugin keeps what it learns on the developer's machine.
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 3 of 7 · from Microsoft Research, 16 Nov 2026
We first profiled Claude Code on 60 tasks from SWE-Bench.
Fact 5 of 7 · from Microsoft Research, 16 Nov 2026
with costs 16.0–21.2% lower on Web-Bench
Fact 6 of 7 · from Microsoft Research, 16 Nov 2026
lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation
Fact 7 of 7 · from Microsoft Research, 16 Nov 2026
keeps what it learns on the developer’s machine
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against this source.
1 TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient
Open the source archived copy-
A Microsoft Research blog post of 16 November 2026 describes TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security that learns from a coding agent's past sessions in a repository to reduce costs in future sessions without modifying the agent or its model.
citeTokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security
-
On SWE-Bench (277 tasks, six repositories, five agent-model combinations) the researchers report mean cost per task fell 18.4% to 30.9%, and pass@1 rose in four of the five combinations.
citeTokenEconomizer lowered the average cost by 18.4–30.9%.
-
The post names Claude Code as the only agent; it does not identify the other four agent-model combinations.
citeWe first profiled Claude Code on 60 tasks from SWE-Bench.
-
The researchers report that in profiling of Claude Code on 60 SWE-Bench tasks, 30% of tool calls repeated an earlier action in the same repository and 16 of 23 tool descriptions sent per turn were never called.
citeThe agent received 23 tool descriptions on every turn, but 16 were never called.
-
The researchers report costs 16.0% to 21.2% lower on Web-Bench and 6.4% to 19.6% lower on security auditing.
citewith costs 16.0–21.2% lower on Web-Bench
-
The researchers say lowering the model's reasoning effort saved similar or larger amounts but substantially degraded performance.
citelowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation
-
The code is at github.com/microsoft/TokenEconomizer, and the post says the plugin keeps what it learns on the developer's machine.
citekeeps what it learns on the developer’s machine
Every sentence above is checked against this source. Each fact has its own address you can cite.
Source 1 · Microsoft Research · 16 Nov 2026
TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient
A Microsoft Research blog post of 16 November 2026 describes TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security that learns from a coding agent's past sessions in a repository to reduce costs in future sessions without modifying the agent or its model.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security
On SWE-Bench (277 tasks, six repositories, five agent-model combinations) the researchers report mean cost per task fell 18.4% to 30.9%, and pass@1 rose in four of the five combinations.
TokenEconomizer lowered the average cost by 18.4–30.9%.
The post names Claude Code as the only agent; it does not identify the other four agent-model combinations.
We first profiled Claude Code on 60 tasks from SWE-Bench.
Show all 7 factsShow fewer facts
The researchers report that in profiling of Claude Code on 60 SWE-Bench tasks, 30% of tool calls repeated an earlier action in the same repository and 16 of 23 tool descriptions sent per turn were never called.
The agent received 23 tool descriptions on every turn, but 16 were never called.
The researchers report costs 16.0% to 21.2% lower on Web-Bench and 6.4% to 19.6% lower on security auditing.
with costs 16.0–21.2% lower on Web-Bench
The researchers say lowering the model's reasoning effort saved similar or larger amounts but substantially degraded performance.
lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation
The code is at github.com/microsoft/TokenEconomizer, and the post says the plugin keeps what it learns on the developer's machine.
keeps what it learns on the developer’s machine
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep